Predominant SARS-CoV-2 variant impacts accuracy when screening for infection using exhaled breath vapor

Abstract

Background: New technologies with novel and ambitious approaches are being developed to diagnose or screen for SARS-CoV-2, including breath tests. The US FDA approved the first breath test for COVID-19 under emergency use authorization in April 2022. Most breath-based assays measure volatile metabolites exhaled by persons to identify a host response to infection. We hypothesized that the breathprint of COVID-19 fluctuated after Omicron became the primary variant of transmission over the Delta variant.

Methods: We collected breath samples from 142 persons with and without a confirmed COVID-19 infection during the Delta and Omicron waves. Breath samples were analyzed by gas chromatography-mass spectrometry.

Results: Here we show that based on 63 exhaled compounds, a general COVID-19 model had an accuracy of 0.73 ± 0.06, which improved to 0.82 ± 0.12 when modeling only the Delta wave, and 0.84 ± 0.06 for the Omicron wave. The specificity improved for the Delta and Omicron models (0.79 ± 0.21 and 0.74 ± 0.12, respectively) relative to the general model (0.61 ± 0.13).

Conclusions: We report that the volatile signature of COVID-19 in breath differs between the Delta-predominant and Omicron-predominant variant waves, and accuracies improve when samples from these waves are modeled separately rather than as one universal approach. Our findings have important implications for groups developing breath-based assays for COVID-19 and other respiratory pathogens, as the host response to infection may significantly differ depending on variants or subtypes.

Fig. 1. Comparisons of breath-based models calibrated with and without regard to COVID variant.

Partial least squares-discriminant analysis (PLS-DA) model comparisons when models were developed using a breath samples collected during the wave of both variants, b just the Delta wave, and c just the Omicron wave. Data show scatter plots of latent variable (LV) scores; boxplots of PLS-DA prediction scores from breath samples with 1 modeled as COVID(−) and 0 as COVID(+); and receiver operator characteristics (ROC) curves showing model accuracies to predict COVID infection from breath volatile organic compounds (VOCs). Analyses were conducted from breath samples of n = 96 COVID(−), n = 12 Delta COVID(+), n = 28 Omicron COVID(+) persons.

Fig. 2. Abundances of exhaled volatile compounds considered by breath-based models to predict COVID-19 infection (see also Fig. 3).

Boxplots representing the abundances of the 63 volatile organic compounds (VOCs) (continued in Fig. 3) used by partial least squares-discriminant analysis (PLS-DA) models to differentiate breath samples of non-COVID (NC) controls, n = 96, from COVID(+) samples collected during the Delta (D), n = 12, and Omicron (O), n = 28, waves. For each compound, the data were normalized to the mean intensity from non-COVID controls, so y-axis values represent the number of times greater relative to the mean non-COVID abundance. Asterisks indicate significant differences per a Kruskal–Wallis test between two groups (*p < 0.05, **p < 0.01). See Supplementary Data 2 to map compound number to putative identification. The central red line indicates median, bottom and top edges indicate 25th and 75th percentiles respectively, whiskers extend to the most extreme non-outlier data points, and outliers are plotted with the red ‘+’ symbol.

Fig. 3. Abundances of exhaled volatile compounds considered by breath-based models to predict COVID-19 infection (see also Fig. 2).

Boxplots representing the abundances of the 63 volatile organic compounds (VOCs) (continued in Fig. 2) used by partial least squares-discriminant analysis (PLS-DA) models to differentiate breath samples of non-COVID (NC) controls, n = 96, from COVID(+) samples collected during the Delta (D), n = 12, and Omicron (O), n = 28, waves. For each compound, the data were normalized to the mean intensity from non-COVID controls, so y-axis values represent the number of times greater relative to the mean non-COVID abundance. Asterisks indicate significant differences per a Kruskal–Wallis test between two groups (*p < 0.05, **p < 0.01). See Supplementary Data 2 to map compound number to putative identification. The central red line indicates median, bottom and top edges indicate 25th and 75th percentiles, respectively, whiskers extend to the most extreme non-outlier data points, and outliers are plotted with the red ‘+’ symbol.